Writing Prompts
Our guidelines on how to write effective prompts
Start Simple
Question-based prompts are the best to start with. Extruct AI is finetuned to automatically determine what actions to take to fulfill your request.
Prompt examples:
What are the recent news associated with company X?
Find careers page URL of company X
What is the mission statement of company X?
Provide “How” Instructions
Some questions are easier to answer than others.
For example, What is the number of employees in company X?
is more straightforward than What is the company X policy against web scraping?
When you are not satisfied with the quality of results, consider instructing Extruct on “how” to find the answer:
- What search queries to search
- What website pages to visit
- What information to ignore
The example above can be improved into:
What is company X policy against web scraping?
Visit the company website and navigate to the "Terms of Service" page to find the answer.
Split Complex Prompts
AI Agents and LLMs often struggle with complex instructions. Thus, we suggest you split your complex task into many smaller ones.
For example, to find & analyze job openings in a company, instead of putting everything into a single task, consider splitting it like this:
- Find the careers page URL.
- Find job postings on the careers page found in Step 1.
- Find job postings of a company on LinkedIn (e.g. using
site:linkedin.com/jobs/view intitle:<company name>
filter) - Analyse job postings from Step 2 and Step 3.
LLMs and autonomous AI agents are rapidly evolving areas of research. The Extruct AI team continuously monitors the best industry practices and cutting-edge research. We believe that AI agents will become increasingly autonomous in a few years as their planning and reasoning capabilities improve.
Customize Output Format
You can directly specify a custom format to the =EXTRUCT()
function in the prompt. We usually add it at the end of the prompt. For example, here is the formatting instructions we use for c_key_people
Prebuilt Agent:
...main prompt...
Output Format: plaintext bulleted list of key people. For every person, print the person's name,
position at the company, and LinkedIn profile (if available) in a format
'- <person name>, <position>: <linkedin profile>'. Avoid any additional commentary or narrative introduction.
Company Context: {input}
Leverage Site Filters
To ask the Agent to search specific websites, tell it to leverage site:
filter. Here is an example from c_linkedin
Prebuilt Agent:
Given the company context, find its LinkedIn profile URL (if it exists). Search internet with
'site:linkedin.com/company' filter. Make sure the profile found belongs to the company provided.
(e.g. matches by company name / website / other context). If you are not sure, output is Not Found.
Company context: {input}
Master Concatenation (&
)
Do not forget spaces before and after the concatenation operator &
. It won’t matter in most cases as the Agent is intelligent enough to understand the question. But it’s a good hygiene practice that might improve the quality of results at no cost.
Bad:
=EXTRUCT("Find the website of the company" & A1 & "visit the website and summarize the landing page")
Good:
=EXTRUCT("Find the website of the company " & A1 & ". Visit the website and summarize the landing page")
Minimize Narrative Introduction
Instruction-tuned LLMs are often trained to repeat the question in the answer, resulting in narrative introductory comments, like Here is a business model description of a company:
.
To minimize this effect, we usually use one of these prompt additions:
Print <data you're looking for> only, nothing else.
Avoid any additional commentary or narrative introduction.
Output <data you're looking for> without any additional commentary.